Highly interpretable classifiers for scikit learn, producing easily understood decision rules instead of black box models
Python Yacc
Latest commit 6f18a42 Apr 20, 2016 Thomas Madl class order fix in predict_proba


Highly interpretable, sklearn-compatible classifier based on decision rules

This is a scikit-learn compatible wrapper for the Bayesian Rule List classifier developed by Letham et al., 2015 (see Letham's original code), extended by a minimum description length-based discretizer (Fayyad & Irani, 1993) for continuous data, and by an approach to subsample large datasets for better performance.

It produces rule lists, which makes trained classifiers easily interpretable to human experts, and is competitive with state of the art classifiers such as random forests or SVMs.

For example, an easily understood Rule List model of the well-known Titanic dataset:

IF male AND adult THEN survival probability: 21% (19% - 23%)
ELSE IF 3rd class THEN survival probability: 44% (38% - 51%)
ELSE IF 1st class THEN survival probability: 96% (92% - 99%)
ELSE survival probability: 88% (82% - 94%)

Letham et al.'s approach only works on discrete data. However, this approach can still be used on continuous data after discretization. The RuleListClassifier class also includes a discretizer that can deal with continuous data (using Fayyad & Irani's minimum description length principle criterion, based on an implementation by navicto).

The inference procedure is slow on large datasets. If you have more than a few thousand data points, and only numeric data, try the included BigDataRuleListClassifier(training_subset=0.1), which first determines a small subset of the training data that is most critical in defining a decision boundary (the data points that are hardest to classify) and learns a rule list only on this subset (you can specify which estimator to use for judging which subset is hardest to classify by passing any sklearn-compatible estimator in the subset_estimator parameter - see examples/diabetes_bigdata_demo.py).


The project requires pyFIM, scikit-learn, and pandas to run.

The included RuleListClassifier works as a scikit-learn estimator, with a model.fit(X,y) method which takes training data X (numpy array or pandas DataFrame; continuous, categorical or mixed data) and labels y.

The learned rules of a trained model can be displayed simply by casting the object as a string, e.g. print model, or by using the model.tostring(decimals=1) method and optionally specifying the rounding precision.

Numerical data in X is automatically discretized. To prevent discretization (e.g. to protect columns containing categorical data represented as integers), pass the list of protected column names in the fit method, e.g. model.fit(X,y,undiscretized_features=['CAT_COLUMN_NAME']) (entries in undiscretized columns will be converted to strings and used as categorical values - see examples/hepatitis_mixeddata_demo.py).

Usage example:

from RuleListClassifier import *
from sklearn.datasets.mldata import fetch_mldata
from sklearn.cross_validation import train_test_split
from sklearn.ensemble import RandomForestClassifier

feature_labels = ["#Pregnant","Glucose concentration test","Blood pressure(mmHg)","Triceps skin fold thickness(mm)","2-Hour serum insulin (mu U/ml)","Body mass index","Diabetes pedigree function","Age (years)"]

data = fetch_mldata("diabetes") # get dataset
y = (data.target+1)/2 # target labels (0 or 1)
Xtrain, Xtest, ytrain, ytest = train_test_split(data.data, y) # split

# train classifier (allow more iterations for better accuracy; use BigDataRuleListClassifier for large datasets)
model = RuleListClassifier(max_iter=10000, class1label="diabetes", verbose=False)
model.fit(Xtrain, ytrain, feature_labels=feature_labels)

print "RuleListClassifier Accuracy:", model.score(Xtest, ytest), "Learned interpretable model:\n", model
print "RandomForestClassifier Accuracy:", RandomForestClassifier().fit(Xtrain, ytrain).score(Xtest, ytest)
RuleListClassifier Accuracy: 0.776041666667 Learned interpretable model:
Trained RuleListClassifier for detecting diabetes
IF Glucose concentration test : 157.5_to_inf THEN probability of diabetes: 81.1% (72.5%-72.5%)
ELSE IF Body mass index : -inf_to_26.3499995 THEN probability of diabetes: 5.2% (1.9%-1.9%)
ELSE IF Glucose concentration test : -inf_to_103.5 THEN probability of diabetes: 14.4% (8.8%-8.8%)
ELSE IF Age (years) : 27.5_to_inf THEN probability of diabetes: 59.6% (51.8%-51.8%)
ELSE IF Glucose concentration test : 103.5_to_127.5 THEN probability of diabetes: 15.9% (8.0%-8.0%)
ELSE probability of diabetes: 44.7% (29.5%-29.5%)

RandomForestClassifier Accuracy: 0.729166666667